8 <object id="TOC" data="graphical-toc.svg" type="image/svg+xml"></object>
10 window.onload=function() { // Wait for the SVG to be loaded before changing it
11 var elem=document.querySelector("#TOC").contentDocument.getElementById("ConfigBox")
12 elem.style="opacity:0.93999999;fill:#ff0000;fill-opacity:0.1;stroke:#000000;stroke-width:0.35277778;stroke-linecap:round;stroke-linejoin:round;stroke-miterlimit:4;stroke-dasharray:none;stroke-dashoffset:0;stroke-opacity:1";
18 A number of options can be given at runtime to change the default
19 SimGrid behavior. For a complete list of all configuration options
20 accepted by the SimGrid version used in your simulator, simply pass
21 the --help configuration flag to your program. If some of the options
22 are not documented on this page, this is a bug that you should please
23 report so that we can fix it. Note that some of the options presented
24 here may not be available in your simulators, depending on the
25 :ref:`compile-time options <install_src_config>` that you used.
27 Setting Configuration Items
28 ---------------------------
30 There is several way to pass configuration options to the simulators.
31 The most common way is to use the ``--cfg`` command line argument. For
32 example, to set the item ``Item`` to the value ``Value``, simply
33 type the following on the command-line:
37 my_simulator --cfg=Item:Value (other arguments)
39 Several ``--cfg`` command line arguments can naturally be used. If you
40 need to include spaces in the argument, don't forget to quote the
41 argument. You can even escape the included quotes (write ``@'`` for ``'`` if
42 you have your argument between simple quotes).
44 Another solution is to use the ``<config>`` tag in the platform file. The
45 only restriction is that this tag must occur before the first
46 platform element (be it ``<zone>``, ``<cluster>``, ``<peer>`` or whatever).
47 The ``<config>`` tag takes an ``id`` attribute, but it is currently
48 ignored so you don't really need to pass it. The important part is that
49 within that tag, you can pass one or several ``<prop>`` tags to specify
50 the configuration to use. For example, setting ``Item`` to ``Value``
51 can be done by adding the following to the beginning of your platform
57 <prop id="Item" value="Value"/>
60 A last solution is to pass your configuration directly in your program
61 with :cpp:func:`simgrid::s4u::Engine::set_config` or :cpp:func:`MSG_config`.
65 #include <simgrid/s4u.hpp>
67 int main(int argc, char *argv[]) {
68 simgrid::s4u::Engine e(&argc, argv);
70 simgrid::s4u::Engine::set_config("Item:Value");
77 Existing Configuration Items
78 ----------------------------
81 The full list can be retrieved by passing ``--help`` and
82 ``--help-cfg`` to an executable that uses SimGrid. Try passing
83 ``help`` as a value to get the list of values accepted by a given
84 option. For example, ``--cfg=plugin:help`` will give you the list
85 of plugins available in your installation of SimGrid.
87 - **contexts/factory:** :ref:`cfg=contexts/factory`
88 - **contexts/guard-size:** :ref:`cfg=contexts/guard-size`
89 - **contexts/nthreads:** :ref:`cfg=contexts/nthreads`
90 - **contexts/stack-size:** :ref:`cfg=contexts/stack-size`
91 - **contexts/synchro:** :ref:`cfg=contexts/synchro`
93 - **cpu/maxmin-selective-update:** :ref:`Cpu Optimization Level <options_model_optim>`
94 - **cpu/model:** :ref:`options_model_select`
95 - **cpu/optim:** :ref:`Cpu Optimization Level <options_model_optim>`
97 - **debug/breakpoint:** :ref:`cfg=debug/breakpoint`
98 - **debug/clean-atexit:** :ref:`cfg=debug/clean-atexit`
99 - **debug/verbose-exit:** :ref:`cfg=debug/verbose-exit`
101 - **exception/cutpath:** :ref:`cfg=exception/cutpath`
103 - **host/model:** :ref:`options_model_select`
105 - **maxmin/precision:** :ref:`cfg=maxmin/precision`
106 - **maxmin/concurrency-limit:** :ref:`cfg=maxmin/concurrency-limit`
108 - **msg/debug-multiple-use:** :ref:`cfg=msg/debug-multiple-use`
110 - **model-check:** :ref:`options_modelchecking`
111 - **model-check/checkpoint:** :ref:`cfg=model-check/checkpoint`
112 - **model-check/communications-determinism:** :ref:`cfg=model-check/communications-determinism`
113 - **model-check/dot-output:** :ref:`cfg=model-check/dot-output`
114 - **model-check/max-depth:** :ref:`cfg=model-check/max-depth`
115 - **model-check/property:** :ref:`cfg=model-check/property`
116 - **model-check/reduction:** :ref:`cfg=model-check/reduction`
117 - **model-check/replay:** :ref:`cfg=model-check/replay`
118 - **model-check/send-determinism:** :ref:`cfg=model-check/send-determinism`
119 - **model-check/termination:** :ref:`cfg=model-check/termination`
120 - **model-check/timeout:** :ref:`cfg=model-check/timeout`
121 - **model-check/visited:** :ref:`cfg=model-check/visited`
123 - **network/bandwidth-factor:** :ref:`cfg=network/bandwidth-factor`
124 - **network/crosstraffic:** :ref:`cfg=network/crosstraffic`
125 - **network/latency-factor:** :ref:`cfg=network/latency-factor`
126 - **network/loopback-lat:** :ref:`cfg=network/loopback`
127 - **network/loopback-bw:** :ref:`cfg=network/loopback`
128 - **network/maxmin-selective-update:** :ref:`Network Optimization Level <options_model_optim>`
129 - **network/model:** :ref:`options_model_select`
130 - **network/optim:** :ref:`Network Optimization Level <options_model_optim>`
131 - **network/TCP-gamma:** :ref:`cfg=network/TCP-gamma`
132 - **network/weight-S:** :ref:`cfg=network/weight-S`
134 - **ns3/TcpModel:** :ref:`options_pls`
135 - **ns3/seed:** :ref:`options_pls`
136 - **path:** :ref:`cfg=path`
137 - **plugin:** :ref:`cfg=plugin`
139 - **storage/max_file_descriptors:** :ref:`cfg=storage/max_file_descriptors`
141 - **surf/precision:** :ref:`cfg=surf/precision`
143 - **For collective operations of SMPI,** please refer to Section :ref:`cfg=smpi/coll-selector`
144 - **smpi/auto-shared-malloc-thresh:** :ref:`cfg=smpi/auto-shared-malloc-thresh`
145 - **smpi/async-small-thresh:** :ref:`cfg=smpi/async-small-thresh`
146 - **smpi/buffering:** :ref:`cfg=smpi/buffering`
147 - **smpi/bw-factor:** :ref:`cfg=smpi/bw-factor`
148 - **smpi/coll-selector:** :ref:`cfg=smpi/coll-selector`
149 - **smpi/comp-adjustment-file:** :ref:`cfg=smpi/comp-adjustment-file`
150 - **smpi/cpu-threshold:** :ref:`cfg=smpi/cpu-threshold`
151 - **smpi/display-allocs:** :ref:`cfg=smpi/display-allocs`
152 - **smpi/display-timing:** :ref:`cfg=smpi/display-timing`
153 - **smpi/errors-are-fatal:** :ref:`cfg=smpi/errors-are-fatal`
154 - **smpi/finalization-barrier:** :ref:`cfg=smpi/finalization-barrier`
155 - **smpi/grow-injected-times:** :ref:`cfg=smpi/grow-injected-times`
156 - **smpi/host-speed:** :ref:`cfg=smpi/host-speed`
157 - **smpi/IB-penalty-factors:** :ref:`cfg=smpi/IB-penalty-factors`
158 - **smpi/iprobe:** :ref:`cfg=smpi/iprobe`
159 - **smpi/iprobe-cpu-usage:** :ref:`cfg=smpi/iprobe-cpu-usage`
160 - **smpi/init:** :ref:`cfg=smpi/init`
161 - **smpi/keep-temps:** :ref:`cfg=smpi/keep-temps`
162 - **smpi/lat-factor:** :ref:`cfg=smpi/lat-factor`
163 - **smpi/ois:** :ref:`cfg=smpi/ois`
164 - **smpi/or:** :ref:`cfg=smpi/or`
165 - **smpi/os:** :ref:`cfg=smpi/os`
166 - **smpi/papi-events:** :ref:`cfg=smpi/papi-events`
167 - **smpi/pedantic:** :ref:`cfg=smpi/pedantic`
168 - **smpi/privatization:** :ref:`cfg=smpi/privatization`
169 - **smpi/privatize-libs:** :ref:`cfg=smpi/privatize-libs`
170 - **smpi/send-is-detached-thresh:** :ref:`cfg=smpi/send-is-detached-thresh`
171 - **smpi/shared-malloc:** :ref:`cfg=smpi/shared-malloc`
172 - **smpi/shared-malloc-hugepage:** :ref:`cfg=smpi/shared-malloc-hugepage`
173 - **smpi/simulate-computation:** :ref:`cfg=smpi/simulate-computation`
174 - **smpi/test:** :ref:`cfg=smpi/test`
175 - **smpi/wtime:** :ref:`cfg=smpi/wtime`
176 - **smpi/list-leaks** :ref:`cfg=smpi/list-leaks`
178 - **Tracing configuration options** can be found in Section :ref:`tracing_tracing_options`
180 - **storage/model:** :ref:`options_model_select`
182 - **vm/model:** :ref:`options_model_select`
186 Configuring the Platform Models
187 -------------------------------
189 .. _options_model_select:
191 Choosing the Platform Models
192 ............................
194 SimGrid comes with several network, CPU and disk models built in,
195 and you can change the used model at runtime by changing the passed
196 configuration. The three main configuration items are given below.
197 For each of these items, passing the special ``help`` value gives you
198 a short description of all possible values (for example,
199 ``--cfg=network/model:help`` will present all provided network
200 models). Also, ``--help-models`` should provide information about all
201 models for all existing resources.
203 - ``network/model``: specify the used network model. Possible values:
205 - **LV08 (default one):** Realistic network analytic model
206 (slow-start modeled by multiplying latency by 13.01, bandwidth by
207 .97; bottleneck sharing uses a payload of S=20537 for evaluating
208 RTT). Described in `Accuracy Study and Improvement of Network
209 Simulation in the SimGrid Framework
210 <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
211 - **Constant:** Simplistic network model where all communication
212 take a constant time (one second). This model provides the lowest
213 realism, but is (marginally) faster.
214 - **SMPI:** Realistic network model specifically tailored for HPC
215 settings (accurate modeling of slow start with correction factors on
216 three intervals: < 1KiB, < 64 KiB, >= 64 KiB). This model can be
217 :ref:`further configured <options_model_network>`.
218 - **IB:** Realistic network model specifically tailored for HPC
219 settings with InfiniBand networks (accurate modeling contention
220 behavior, based on the model explained in `this PhD work
221 <http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_.
222 This model can be :ref:`further configured <options_model_network>`.
223 - **CM02:** Legacy network analytic model. Very similar to LV08, but
224 without corrective factors. The timings of small messages are thus
225 poorly modeled. This model is described in `A Network Model for
226 Simulation of Grid Application
227 <https://hal.inria.fr/inria-00071989/document>`_.
228 - **ns-3** (only available if you compiled SimGrid accordingly):
229 Use the packet-level network
230 simulators as network models (see :ref:`model_ns3`).
231 This model can be :ref:`further configured <options_pls>`.
233 - ``cpu/model``: specify the used CPU model. We have only one model
236 - **Cas01:** Simplistic CPU model (time=size/speed)
238 - ``host/model``: The host concept is the aggregation of a CPU with a
239 network card. Three models exists, but actually, only 2 of them are
240 interesting. The "compound" one is simply due to the way our
241 internal code is organized, and can easily be ignored. So at the
242 end, you have two host models: The default one allows aggregation of
243 an existing CPU model with an existing network model, but does not
244 allow parallel tasks because these beasts need some collaboration
245 between the network and CPU model. That is why, ptask_07 is used by
246 default when using SimDag.
248 - **default:** Default host model. Currently, CPU:Cas01 and
249 network:LV08 (with cross traffic enabled)
250 - **compound:** Host model that is automatically chosen if
251 you change the network and CPU models
252 - **ptask_L07:** Host model somehow similar to Cas01+CM02 but
253 allowing "parallel tasks", that are intended to model the moldable
254 tasks of the grid scheduling literature.
256 - ``storage/model``: specify the used storage model. Only one model is
258 - ``vm/model``: specify the model for virtual machines. Only one model
261 .. todo: make 'compound' the default host model.
263 .. _options_model_optim:
268 The network and CPU models that are based on lmm_solve (that
269 is, all our analytical models) accept specific optimization
272 - items ``network/optim`` and ``cpu/optim`` (both default to 'Lazy'):
274 - **Lazy:** Lazy action management (partial invalidation in lmm +
275 heap in action remaining).
276 - **TI:** Trace integration. Highly optimized mode when using
277 availability traces (only available for the Cas01 CPU model for
279 - **Full:** Full update of remaining and variables. Slow but may be
280 useful when debugging.
282 - items ``network/maxmin-selective-update`` and
283 ``cpu/maxmin-selective-update``: configure whether the underlying
284 should be lazily updated or not. It should have no impact on the
285 computed timings, but should speed up the computation. |br| It is
286 still possible to disable this feature because it can reveal
287 counter-productive in very specific scenarios where the
288 interaction level is high. In particular, if all your
289 communication share a given backbone link, you should disable it:
290 without it, a simple regular loop is used to update each
291 communication. With it, each of them is still updated (because of
292 the dependency induced by the backbone), but through a complicated
293 and slow pattern that follows the actual dependencies.
295 .. _cfg=maxmin/precision:
296 .. _cfg=surf/precision:
301 **Option** ``maxmin/precision`` **Default:** 0.00001 (in flops or bytes) |br|
302 **Option** ``surf/precision`` **Default:** 0.00001 (in seconds)
304 The analytical models handle a lot of floating point values. It is
305 possible to change the epsilon used to update and compare them through
306 this configuration item. Changing it may speedup the simulation by
307 discarding very small actions, at the price of a reduced numerical
308 precision. You can modify separately the precision used to manipulate
309 timings (in seconds) and the one used to manipulate amounts of work
312 .. _cfg=maxmin/concurrency-limit:
317 **Option** ``maxmin/concurrency-limit`` **Default:** -1 (no limit)
319 The maximum number of variables per resource can be tuned through this
320 option. You can have as many simultaneous actions per resources as you
321 want. If your simulation presents a very high level of concurrency, it
322 may help to use e.g. 100 as a value here. It means that at most 100
323 actions can consume a resource at a given time. The extraneous actions
324 are queued and wait until the amount of concurrency of the considered
325 resource lowers under the given boundary.
327 Such limitations help both to the simulation speed and simulation accuracy
328 on highly constrained scenarios, but the simulation speed suffers of this
329 setting on regular (less constrained) scenarios so it is off by default.
331 .. _options_model_network:
333 Configuring the Network Model
334 .............................
336 .. _cfg=network/TCP-gamma:
338 Maximal TCP Window Size
339 ^^^^^^^^^^^^^^^^^^^^^^^
341 **Option** ``network/TCP-gamma`` **Default:** 4194304
343 The analytical models need to know the maximal TCP window size to take
344 the TCP congestion mechanism into account. On Linux, this value can
345 be retrieved using the following commands. Both give a set of values,
346 and you should use the last one, which is the maximal size.
348 .. code-block:: shell
350 cat /proc/sys/net/ipv4/tcp_rmem # gives the sender window
351 cat /proc/sys/net/ipv4/tcp_wmem # gives the receiver window
353 .. _cfg=network/bandwidth-factor:
354 .. _cfg=network/latency-factor:
355 .. _cfg=network/weight-S:
357 Correcting Important Network Parameters
358 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
360 SimGrid can take network irregularities such as a slow startup or
361 changing behavior depending on the message size into account. You
362 should not change these values unless you really know what you're
363 doing. The corresponding values were computed through data fitting
364 one the timings of packet-level simulators, as described in `Accuracy
365 Study and Improvement of Network Simulation in the SimGrid Framework
366 <http://mescal.imag.fr/membres/arnaud.legrand/articles/simutools09.pdf>`_.
369 If you are using the SMPI model, these correction coefficients are
370 themselves corrected by constant values depending on the size of the
371 exchange. By default SMPI uses factors computed on the Stampede
372 Supercomputer at TACC, with optimal deployment of processes on
373 nodes. Again, only hardcore experts should bother about this fact.
376 .. todo:: This section should be rewritten, and actually explain the
377 options network/bandwidth-factor, network/latency-factor,
380 .. _cfg=smpi/IB-penalty-factors:
385 InfiniBand network behavior can be modeled through 3 parameters
386 ``smpi/IB-penalty-factors:"βe;βs;γs"``, as explained in `this PhD
388 <http://mescal.imag.fr/membres/jean-marc.vincent/index.html/PhD/Vienne.pdf>`_ (in French)
389 or more concisely in `this paper <https://hal.inria.fr/hal-00953618/document>`_,
390 even if that paper does only describe models for myrinet and ethernet.
391 You can see in Fig 2 some results for Infiniband, for example. This model
392 may be outdated by now for modern infiniband, anyway, so a new
393 validation would be good.
395 The three paramaters are defined as follows:
397 - βs: penalty factor for outgoing messages, computed by running a simple send to
398 two nodes and checking slowdown compared to a single send to one node,
400 - βe: penalty factor for ingoing messages, same computation method but with one
401 node receiving several messages
402 - γr: slowdown factor when communication buffer memory is saturated. It needs a
403 more complicated pattern to run in order to be computed (5.3 in the thesis,
404 page 107), and formula in the end is γr = time(c)/(3×βe×time(ref)), where
405 time(ref) is the time of a single comm with no contention).
407 Once these values are computed, a penalty is assessed for each message (this is
408 the part implemented in the simulator) as shown page 106 of the thesis. Here is
409 a simple translation of this text. First, some notations:
411 - ∆e(e) which corresponds to the incoming degree of node e, that is to say the number of communications having as destination node e.
412 - ∆s (s) which corresponds to the degree outgoing from node s, that is to say the number of communications sent by node s.
413 - Φ (e) which corresponds to the number of communications destined for the node e but coming from a different node.
414 - Ω (s, e) which corresponds to the number of messages coming from node s to node e. If node e only receives communications from different nodes then Φ (e) = ∆e (e). On the other hand if, for example, there are three messages coming from node s and going from node e then Φ (e) 6 = ∆e (e) and Ω (s, e) = 3
416 To determine the penalty for a communication, two values need to be calculated. First, the penalty caused by the conflict in transmission, noted ps.
419 - if ∆s (i) = 1 then ps = 1.
420 - if ∆s (i) ≥ 2 and ∆e (i) ≥ 3 then ps = ∆s (i) × βs × γr
421 - else, ps = ∆s (i) × βs
424 Then, the penalty caused by the conflict in reception (noted pe) should be computed as follows:
426 - if ∆e (i) = 1 then pe = 1
427 - else, pe = Φ (e) × βe × Ω (s, e)
429 Finally, the penalty associated with the communication is:
430 p = max (ps ∈ s, pe)
432 .. _cfg=network/crosstraffic:
434 Simulating Cross-Traffic
435 ^^^^^^^^^^^^^^^^^^^^^^^^
437 Since SimGrid v3.7, cross-traffic effects can be taken into account in
438 analytical simulations. It means that ongoing and incoming
439 communication flows are treated independently. In addition, the LV08
440 model adds 0.05 of usage on the opposite direction for each new
441 created flow. This can be useful to simulate some important TCP
442 phenomena such as ack compression.
444 For that to work, your platform must have two links for each
445 pair of interconnected hosts. An example of usable platform is
446 available in ``examples/platforms/crosstraffic.xml``.
448 This is activated through the ``network/crosstraffic`` item, that
449 can be set to 0 (disable this feature) or 1 (enable it).
451 Note that with the default host model this option is activated by default.
453 .. _cfg=network/loopback:
455 Configuring loopback link
456 ^^^^^^^^^^^^^^^^^^^^^^^^^
458 Several network model provide an implicit loopback link to account for local
459 communication on a host. By default it has a 10GBps bandwidth and a null latency.
460 This can be changed with ``network/loopback-lat`` and ``network/loopback-bw``
463 .. _cfg=smpi/async-small-thresh:
465 Simulating Asynchronous Send
466 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^
468 (this configuration item is experimental and may change or disappear)
470 It is possible to specify that messages below a certain size (in bytes) will be
471 sent as soon as the call to MPI_Send is issued, without waiting for
472 the correspondent receive. This threshold can be configured through
473 the ``smpi/async-small-thresh`` item. The default value is 0. This
474 behavior can also be manually set for mailboxes, by setting the
475 receiving mode of the mailbox with a call to
476 :cpp:func:`MSG_mailbox_set_async`. After this, all messages sent to
477 this mailbox will have this behavior regardless of the message size.
479 This value needs to be smaller than or equals to the threshold set at
480 :ref:`cfg=smpi/send-is-detached-thresh`, because asynchronous messages
481 are meant to be detached as well.
488 **Option** ``ns3/TcpModel`` **Default:** "default" (ns-3 default)
490 When using ns-3, there is an extra item ``ns3/TcpModel``, corresponding
491 to the ``ns3::TcpL4Protocol::SocketType`` configuration item in
492 ns-3. The only valid values (enforced on the SimGrid side) are
493 'default' (no change to the ns-3 configuration), 'NewReno' or 'Reno' or
496 **Option** ``ns3/seed`` **Default:** "" (don't set the seed in ns-3)
498 This option is the random seed to provide to ns-3 with
499 ``ns3::RngSeedManager::SetSeed`` and ``ns3::RngSeedManager::SetRun``.
501 If left blank, no seed is set in ns-3. If the value 'time' is
502 provided, the current amount of seconds since epoch is used as a seed.
503 Otherwise, the provided value must be a number to use as a seed.
505 Configuring the Storage model
506 .............................
508 .. _cfg=storage/max_file_descriptors:
510 File Descriptor Count per Host
511 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
513 **Option** ``storage/max_file_descriptors`` **Default:** 1024
515 Each host maintains a fixed-size array of its file descriptors. You
516 can change its size through this item to either enlarge it if your
517 application requires it or to reduce it to save memory space.
524 SimGrid plugins allow one to extend the framework without changing its
525 source code directly. Read the source code of the existing plugins to
526 learn how to do so (in ``src/plugins``), and ask your questions to the
527 usual channels (Stack Overflow, Mailing list, IRC). The basic idea is
528 that plugins usually register callbacks to some signals of interest.
529 If they need to store some information about a given object (Link, CPU
530 or Actor), they do so through the use of a dedicated object extension.
532 Some of the existing plugins can be activated from the command line,
533 meaning that you can activate them from the command line without any
534 modification to your simulation code. For example, you can activate
535 the host energy plugin by adding ``--cfg=plugin:host_energy`` to your
538 Here is a partial list of plugins that can be activated this way. You can get
539 the full list by passing ``--cfg=plugin:help`` to your simulator.
541 - :ref:`Host Energy <plugin_host_energy>`: models the energy dissipation of the compute units.
542 - :ref:`Link Energy <plugin_link_energy>`: models the energy dissipation of the network.
543 - :ref:`Host Load <plugin_host_load>`: monitors the load of the compute units.
545 .. _options_modelchecking:
547 Configuring the Model-Checking
548 ------------------------------
550 To enable SimGrid's model-checking support, the program should
551 be executed using the simgrid-mc wrapper:
553 .. code-block:: shell
555 simgrid-mc ./my_program
557 Safety properties are expressed as assertions using the function
558 :cpp:func:`void MC_assert(int prop)`.
560 .. _cfg=smpi/buffering:
562 Specifying the MPI buffering behavior
563 .....................................
565 **Option** ``smpi/buffering`` **Default:** infty
567 Buffering in MPI has a huge impact on the communication semantic. For example,
568 standard blocking sends are synchronous calls when the system buffers are full
569 while these calls can complete immediately without even requiring a matching
570 receive call for small messages sent when the system buffers are empty.
572 In SMPI, this depends on the message size, that is compared against two thresholds:
574 - if (size < :ref:`smpi/async-small-thresh <cfg=smpi/async-small-thresh>`) then
575 MPI_Send returns immediately, even if the corresponding receive has not be issued yet.
576 - if (:ref:`smpi/async-small-thresh <cfg=smpi/async-small-thresh>` < size < :ref:`smpi/send-is-detached-thresh <cfg=smpi/send-is-detached-thresh>`) then
577 MPI_Send returns as soon as the corresponding receive has been issued. This is known as the eager mode.
578 - if (:ref:`smpi/send-is-detached-thresh <cfg=smpi/send-is-detached-thresh>` < size) then
579 MPI_Send returns only when the message has actually been sent over the network. This is known as the rendez-vous mode.
581 The ``smpi/buffering`` (only valid with MC) option gives an easier interface to choose between these semantics. It can take two values:
583 - **zero:** means that buffering should be disabled. All communications are actually blocking.
584 - **infty:** means that buffering should be made infinite. All communications are non-blocking.
586 .. _cfg=model-check/property:
588 Specifying a liveness property
589 ..............................
591 **Option** ``model-check/property`` **Default:** unset
593 If you want to specify liveness properties, you have to pass them on
594 the command line, specifying the name of the file containing the
595 property, as formatted by the `ltl2ba <https://github.com/utwente-fmt/ltl2ba>`_ program.
596 Note that ltl2ba is not part of SimGrid and must be installed separately.
598 .. code-block:: shell
600 simgrid-mc ./my_program --cfg=model-check/property:<filename>
602 .. _cfg=model-check/checkpoint:
604 Going for Stateful Verification
605 ...............................
607 By default, the system is backtracked to its initial state to explore
608 another path, instead of backtracking to the exact step before the fork
609 that we want to explore (this is called stateless verification). This
610 is done this way because saving intermediate states can rapidly
611 exhaust the available memory. If you want, you can change the value of
612 the ``model-check/checkpoint`` item. For example,
613 ``--cfg=model-check/checkpoint:1`` asks to take a checkpoint every
614 step. Beware, this will certainly explode your memory. Larger values
615 are probably better, make sure to experiment a bit to find the right
616 setting for your specific system.
618 .. _cfg=model-check/reduction:
620 Specifying the kind of reduction
621 ................................
623 The main issue when using the model-checking is the state space
624 explosion. You can activate some reduction technique with
625 ``--cfg=model-check/reduction:<technique>``. For now, this
626 configuration variable can take 2 values:
628 - **none:** Do not apply any kind of reduction (mandatory for
629 liveness properties, as our current DPOR algorithm breaks cycles)
630 - **dpor:** Apply Dynamic Partial Ordering Reduction. Only valid if
631 you verify local safety properties (default value for safety
634 Another way to mitigate the state space explosion is to search for
635 cycles in the exploration with the :ref:`cfg=model-check/visited`
636 configuration. Note that DPOR and state-equality reduction may not
637 play well together. You should choose between them.
639 Our current DPOR implementation could be improved in may ways. We are
640 currently improving its efficiency (both in term of reduction ability
641 and computational speed), and future work could make it compatible
642 with liveness properties.
644 .. _cfg=model-check/visited:
646 Size of Cycle Detection Set (state equality reduction)
647 ......................................................
649 Mc SimGrid can be asked to search for cycles during the exploration,
650 i.e. situations where a new explored state is in fact the same state
651 than a previous one.. This can prove useful to mitigate the state
652 space explosion with safety properties, and this is the crux when
653 searching for counter-examples to the liveness properties.
655 Note that this feature may break the current implementation of the
656 DPOR reduction technique.
658 The ``model-check/visited`` item is the maximum number of states, which
659 are stored in memory. If the maximum number of snapshotted state is
660 reached, some states will be removed from the memory and some cycles
661 might be missed. Small values can lead to incorrect verifications, but
662 large values can exhaust your memory and be CPU intensive as each new
663 state must be compared to that amount of older saved states.
665 The default settings depend on the kind of exploration. With safety
666 checking, no state is snapshotted and cycles cannot be detected. With
667 liveness checking, all states are snapshotted because missing a cycle
668 could hinder the exploration soundness.
670 .. _cfg=model-check/termination:
672 Non-Termination Detection
673 .........................
675 The ``model-check/termination`` configuration item can be used to
676 report if a non-termination execution path has been found. This is a
677 path with a cycle, which means that the program might never terminate.
679 This only works in safety mode, not in liveness mode.
681 This options is disabled by default.
683 .. _cfg=model-check/dot-output:
688 If set, the ``model-check/dot-output`` configuration item is the name
689 of a file in which to write a dot file of the path leading to the
690 property violation discovered (safety or liveness violation), as well
691 as the cycle for liveness properties. This dot file can then be fed to the
692 graphviz dot tool to generate a corresponding graphical representation.
694 .. _cfg=model-check/max-depth:
696 Exploration Depth Limit
697 .......................
699 The ``model-check/max-depth`` can set the maximum depth of the
700 exploration graph of the model checker. If this limit is reached, a
701 logging message is sent and the results might not be exact.
703 By default, the exploration is limited to the depth of 1000.
705 .. _cfg=model-check/timeout:
710 By default, the model checker does not handle timeout conditions: the `wait`
711 operations never time out. With the ``model-check/timeout`` configuration item
712 set to **yes**, the model checker will explore timeouts of `wait` operations.
714 .. _cfg=model-check/communications-determinism:
715 .. _cfg=model-check/send-determinism:
717 Communication Determinism
718 .........................
720 The ``model-check/communications-determinism`` and
721 ``model-check/send-determinism`` items can be used to select the
722 communication determinism mode of the model checker, which checks
723 determinism properties of the communications of an application.
727 Verification Performance Considerations
728 .......................................
730 The size of the stacks can have a huge impact on the memory
731 consumption when using model-checking. By default, each snapshot will
732 save a copy of the whole stacks and not only of the part that is
733 really meaningful: you should expect the contribution of the memory
734 consumption of the snapshots to be:
735 :math:`\text{number of processes} \times \text{stack size} \times \text{number of states}`.
737 When compiled against the model checker, the stacks are not
738 protected with guards: if the stack size is too small for your
739 application, the stack will silently overflow into other parts of the
740 memory (see :ref:`contexts/guard-size <cfg=contexts/guard-size>`).
742 .. _cfg=model-check/replay:
744 Replaying buggy execution paths from the model checker
745 ......................................................
747 Debugging the problems reported by the model checker is challenging:
748 First, the application under verification cannot be debugged with gdb
749 because the model checker already traces it. Then, the model checker may
750 explore several execution paths before encountering the issue, making it
751 very difficult to understand the output. Fortunately, SimGrid provides
752 the execution path leading to any reported issue so that you can replay
753 this path reported by the model checker, enabling the usage of classical
756 When the model checker finds an interesting path in the application
757 execution graph (where a safety or liveness property is violated), it
758 generates an identifier for this path. Here is an example of the output:
760 .. code-block:: shell
762 [ 0.000000] (0:@) Check a safety property
763 [ 0.000000] (0:@) **************************
764 [ 0.000000] (0:@) *** PROPERTY NOT VALID ***
765 [ 0.000000] (0:@) **************************
766 [ 0.000000] (0:@) Counter-example execution trace:
767 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(3)
768 [ 0.000000] (0:@) [(1)Tremblay (app)] MC_RANDOM(4)
769 [ 0.000000] (0:@) Path = 1/3;1/4
770 [ 0.000000] (0:@) Expanded states = 27
771 [ 0.000000] (0:@) Visited states = 68
772 [ 0.000000] (0:@) Executed transitions = 46
774 The interesting line is ``Path = 1/3;1/4``, which means that you should use
775 ``--cfg=model-check/replay:1/3;1/4`` to replay your application on the buggy
776 execution path. All options (but the model checker related ones) must
777 remain the same. In particular, if you ran your application with
778 ``smpirun -wrapper simgrid-mc``, then do it again. Remove all
779 MC-related options, keep non-MC-related ones and add
780 ``--cfg=model-check/replay:???``.
782 Currently, if the path is of the form ``X;Y;Z``, each number denotes
783 the actor's pid that is selected at each indecision point. If it's of
784 the form ``X/a;Y/b``, the X and Y are the selected pids while the a
785 and b are the return values of their simcalls. In the previous
786 example, ``1/3;1/4``, you can see from the full output that the actor
787 1 is doing MC_RANDOM simcalls, so the 3 and 4 simply denote the values
788 that these simcall return on the execution branch leading to the
791 Configuring the User Code Virtualization
792 ----------------------------------------
794 .. _cfg=contexts/factory:
796 Selecting the Virtualization Factory
797 ....................................
799 **Option** contexts/factory **Default:** "raw"
801 In SimGrid, the user code is virtualized in a specific mechanism that
802 allows the simulation kernel to control its execution: when a user
803 process requires a blocking action (such as sending a message), it is
804 interrupted, and only gets released when the simulated clock reaches
805 the point where the blocking operation is done. This is explained
806 graphically in the `relevant tutorial, available online
807 <https://simgrid.org/tutorials/simgrid-simix-101.pdf>`_.
809 In SimGrid, the containers in which user processes are virtualized are
810 called contexts. Several context factory are provided, and you can
811 select the one you want to use with the ``contexts/factory``
812 configuration item. Some of the following may not exist on your
813 machine because of portability issues. In any case, the default one
814 should be the most effcient one (please report bugs if the
815 auto-detection fails for you). They are approximately sorted here from
816 the slowest to the most efficient:
818 - **thread:** very slow factory using full featured threads (either
819 pthreads or windows native threads). They are slow but very
820 standard. Some debuggers or profilers only work with this factory.
821 - **java:** Java applications are virtualized onto java threads (that
822 are regular pthreads registered to the JVM)
823 - **ucontext:** fast factory using System V contexts (Linux and FreeBSD only)
824 - **boost:** This uses the `context
825 implementation <http://www.boost.org/doc/libs/1_59_0/libs/context/doc/html/index.html>`_
826 of the boost library for a performance that is comparable to our
828 |br| Install the relevant library (e.g. with the
829 libboost-contexts-dev package on Debian/Ubuntu) and recompile
831 - **raw:** amazingly fast factory using a context switching mechanism
832 of our own, directly implemented in assembly (only available for x86
833 and amd64 platforms for now) and without any unneeded system call.
835 The main reason to change this setting is when the debugging tools become
836 fooled by the optimized context factories. Threads are the most
837 debugging-friendly contexts, as they allow one to set breakpoints
838 anywhere with gdb and visualize backtraces for all processes, in order
839 to debug concurrency issues. Valgrind is also more comfortable with
840 threads, but it should be usable with all factories (Exception: the
841 callgrind tool really dislikes raw and ucontext factories).
843 .. _cfg=contexts/stack-size:
845 Adapting the Stack Size
846 .......................
848 **Option** ``contexts/stack-size`` **Default:** 8192 KiB
850 Each virtualized used process is executed using a specific system
851 stack. The size of this stack has a huge impact on the simulation
852 scalability, but its default value is rather large. This is because
853 the error messages that you get when the stack size is too small are
854 rather disturbing: this leads to stack overflow (overwriting other
855 stacks), leading to segfaults with corrupted stack traces.
857 If you want to push the scalability limits of your code, you might
858 want to reduce the ``contexts/stack-size`` item. Its default value is
859 8192 (in KiB), while our Chord simulation works with stacks as small
860 as 16 KiB, for example. You can ensure that some actors have a specific
861 size by simply changing the value of this configuration item before
862 creating these actors. The :cpp:func:`simgrid::s4u::Engine::set_config`
863 functions are handy for that.
865 This *setting is ignored* when using the thread factory (because there
866 is no way to modify the stack size with C++ system threads). Instead,
867 you should compile SimGrid and your application with
868 ``-fsplit-stack``. Note that this compilation flag is not compatible
869 with the model checker right now.
871 The operating system should only allocate memory for the pages of the
872 stack which are actually used and you might not need to use this in
873 most cases. However, this setting is very important when using the
874 model checker (see :ref:`options_mc_perf`).
876 .. _cfg=contexts/guard-size:
878 Disabling Stack Guard Pages
879 ...........................
881 **Option** ``contexts/guard-size`` **Default** 1 page in most case (0 pages on Windows or with MC)
883 Unless you use the threads context factory (see
884 :ref:`cfg=contexts/factory`), a stack guard page is usually used
885 which prevents the stack of a given actor from overflowing on another
886 stack. But the performance impact may become prohibitive when the
887 amount of actors increases. The option ``contexts/guard-size`` is the
888 number of stack guard pages used. By setting it to 0, no guard pages
889 will be used: in this case, you should avoid using small stacks (with
890 :ref:`contexts/stack-size <cfg=contexts/stack-size>`) as the stack
891 will silently overflow on other parts of the memory.
893 When no stack guard page is created, stacks may then silently overflow
894 on other parts of the memory if their size is too small for the
897 .. _cfg=contexts/nthreads:
898 .. _cfg=contexts/synchro:
900 Running User Code in Parallel
901 .............................
903 Parallel execution of the user code is only considered stable in
904 SimGrid v3.7 and higher, and mostly for MSG simulations. SMPI
905 simulations may well fail in parallel mode. It is described in
906 `INRIA RR-7653 <http://hal.inria.fr/inria-00602216/>`_.
908 If you are using the **ucontext** or **raw** context factories, you can
909 request to execute the user code in parallel. Several threads are
910 launched, each of them handling the same number of user contexts at each
911 run. To activate this, set the ``contexts/nthreads`` item to the amount
912 of cores that you have in your computer (or lower than 1 to have the
913 amount of cores auto-detected).
915 When parallel execution is activated, you can choose the
916 synchronization schema used with the ``contexts/synchro`` item,
917 which value is either:
919 - **futex:** ultra optimized synchronisation schema, based on futexes
920 (fast user-mode mutexes), and thus only available on Linux systems.
921 This is the default mode when available.
922 - **posix:** slow but portable synchronisation using only POSIX
924 - **busy_wait:** not really a synchronisation: the worker threads
925 constantly request new contexts to execute. It should be the most
926 efficient synchronisation schema, but it loads all the cores of
927 your machine for no good reason. You probably prefer the other less
930 Configuring the Tracing
931 -----------------------
933 The :ref:`tracing subsystem <outcomes_vizu>` can be configured in
934 several different ways depending on the used interface (S4U, SMPI, SimDag)
935 and the kind of traces that needs to be obtained. See the
936 :ref:`Tracing Configuration Options subsection
937 <tracing_tracing_options>` for a full description of each
938 configuration option.
940 We detail here a simple way to get the traces working for you, even if
941 you never used the tracing API.
944 - Any SimGrid-based simulator (MSG, SimDag, SMPI, ...) and raw traces:
946 .. code-block:: shell
948 --cfg=tracing:yes --cfg=tracing/uncategorized:yes
950 The first parameter activates the tracing subsystem, and the second
951 tells it to trace host and link utilization (without any
954 - MSG or SimDag-based simulator and categorized traces (you need to
955 declare categories and classify your tasks according to them)
957 .. code-block:: shell
959 --cfg=tracing:yes --cfg=tracing/categorized:yes
961 The first parameter activates the tracing subsystem, and the second
962 tells it to trace host and link categorized utilization.
964 - SMPI simulator and traces for a space/time view:
966 .. code-block:: shell
970 The `-trace` parameter for the smpirun script runs the simulation
971 with ``--cfg=tracing:yes --cfg=tracing/smpi:yes``. Check the
972 smpirun's `-help` parameter for additional tracing options.
974 Sometimes you might want to put additional information on the trace to
975 correctly identify them later, or to provide data that can be used to
976 reproduce an experiment. You have two ways to do that:
978 - Add a string on top of the trace file as comment:
980 .. code-block:: shell
982 --cfg=tracing/comment:my_simulation_identifier
984 - Add the contents of a textual file on top of the trace file as comment:
986 .. code-block:: shell
988 --cfg=tracing/comment-file:my_file_with_additional_information.txt
990 Please, use these two parameters (for comments) to make reproducible
991 simulations. For additional details about this and all tracing
992 options, check See the :ref:`tracing_tracing_options`.
997 .. _cfg=msg/debug-multiple-use:
1002 **Option** ``msg/debug-multiple-use`` **Default:** off
1004 Sometimes your application may try to send a task that is still being
1005 executed somewhere else, making it impossible to send this task. However,
1006 for debugging purposes, one may want to know what the other host is/was
1007 doing. This option shows a backtrace of the other process.
1012 The SMPI interface provides several specific configuration items.
1013 These are not easy to see, since the code is usually launched through the
1014 ``smiprun`` script directly.
1016 .. _cfg=smpi/host-speed:
1017 .. _cfg=smpi/cpu-threshold:
1018 .. _cfg=smpi/simulate-computation:
1020 Automatic Benchmarking of SMPI Code
1021 ...................................
1023 In SMPI, the sequential code is automatically benchmarked, and these
1024 computations are automatically reported to the simulator. That is to
1025 say that if you have a large computation between a ``MPI_Recv()`` and
1026 a ``MPI_Send()``, SMPI will automatically benchmark the duration of
1027 this code, and create an execution task within the simulator to take
1028 this into account. For that, the actual duration is measured on the
1029 host machine and then scaled to the power of the corresponding
1030 simulated machine. The variable ``smpi/host-speed`` allows one to
1031 specify the computational speed of the host machine (in flop/s by
1032 default) to use when scaling the execution times.
1034 The default value is ``smpi/host-speed=20kf`` (= 20,000 flop/s). This
1035 is probably underestimated for most machines, leading SimGrid to
1036 overestimate the amount of flops in the execution blocks that are
1037 automatically injected in the simulator. As a result, the execution
1038 time of the whole application will probably be overestimated until you
1039 use a realistic value.
1041 When the code consists of numerous consecutive MPI calls, the
1042 previous mechanism feeds the simulation kernel with numerous tiny
1043 computations. The ``smpi/cpu-threshold`` item becomes handy when this
1044 impacts badly on the simulation performance. It specifies a threshold (in
1045 seconds) below which the execution chunks are not reported to the
1046 simulation kernel (default value: 1e-6).
1048 .. note:: The option ``smpi/cpu-threshold`` ignores any computation
1049 time spent below this threshold. SMPI does not consider the
1050 `amount of time` of these computations; there is no offset for
1051 this. Hence, a value that is too small, may lead to unreliable
1054 In some cases, however, one may wish to disable simulation of
1055 the computation of an application. This is the case when SMPI is used not to
1056 simulate an MPI application, but instead an MPI code that performs
1057 "live replay" of another MPI app (e.g., ScalaTrace's replay tool, or
1058 various on-line simulators that run an app at scale). In this case the
1059 computation of the replay/simulation logic should not be simulated by
1060 SMPI. Instead, the replay tool or on-line simulator will issue
1061 "computation events", which correspond to the actual MPI simulation
1062 being replayed/simulated. At the moment, these computation events can
1063 be simulated using SMPI by calling internal smpi_execute*() functions.
1065 To disable the benchmarking/simulation of a computation in the simulated
1066 application, the variable ``smpi/simulate-computation`` should be set
1067 to **no**. This option just ignores the timings in your simulation; it
1068 still executes the computations itself. If you want to stop SMPI from
1069 doing that, you should check the SMPI_SAMPLE macros, documented in
1070 Section :ref:`SMPI_use_faster`.
1072 +------------------------------------+-------------------------+-----------------------------+
1073 | Solution | Computations executed? | Computations simulated? |
1074 +====================================+=========================+=============================+
1075 | --cfg=smpi/simulate-computation:no | Yes | Never |
1076 +------------------------------------+-------------------------+-----------------------------+
1077 | --cfg=smpi/cpu-threshold:42 | Yes, in all cases | If it lasts over 42 seconds |
1078 +------------------------------------+-------------------------+-----------------------------+
1079 | SMPI_SAMPLE() macro | Only once per loop nest | Always |
1080 +------------------------------------+-------------------------+-----------------------------+
1082 .. _cfg=smpi/comp-adjustment-file:
1084 Slow-down or speed-up parts of your code
1085 ........................................
1087 **Option** ``smpi/comp-adjustment-file:`` **Default:** unset
1089 This option allows you to pass a file that contains two columns: The
1090 first column defines the section that will be subject to a speedup;
1091 the second column is the speedup. For instance:
1093 .. code-block:: shell
1095 "start:stop","ratio"
1096 "exchange_1.f:30:exchange_1.f:130",1.18244559422142
1098 The first line is the header - you must include it. The following
1099 line means that the code between two consecutive MPI calls on line 30
1100 in exchange_1.f and line 130 in exchange_1.f should receive a speedup
1101 of 1.18244559422142. The value for the second column is therefore a
1102 speedup, if it is larger than 1 and a slowdown if it is smaller
1103 than 1. Nothing will be changed if it is equal to 1.
1105 Of course, you can set any arbitrary filenames you want (so the start
1106 and end don't have to be in the same file), but be aware that this
1107 mechanism only supports `consecutive calls!`
1109 Please note that you must pass the ``-trace-call-location`` flag to
1110 smpicc or smpiff, respectively. This flag activates some internal
1111 macro definitions that help with obtaining the call location.
1113 .. _cfg=smpi/bw-factor:
1118 **Option** ``smpi/bw-factor``
1119 |br| **Default:** 65472:0.940694;15424:0.697866;9376:0.58729;5776:1.08739;3484:0.77493;1426:0.608902;732:0.341987;257:0.338112;0:0.812084
1121 The possible throughput of network links is often dependent on the
1122 message sizes, as protocols may adapt to different message sizes. With
1123 this option, a series of message sizes and factors are given, helping
1124 the simulation to be more realistic. For instance, the current default
1125 value means that messages with size 65472 bytes and more will get a total of
1126 MAX_BANDWIDTH*0.940694, messages of size 15424 to 65471 will get
1127 MAX_BANDWIDTH*0.697866, and so on (where MAX_BANDWIDTH denotes the
1128 bandwidth of the link).
1130 An experimental script to compute these factors is available online. See
1131 https://framagit.org/simgrid/platform-calibration/
1132 https://simgrid.org/contrib/smpi-saturation-doc.html
1134 .. _cfg=smpi/display-timing:
1136 Reporting Simulation Time
1137 .........................
1139 **Option** ``smpi/display-timing`` **Default:** 0 (false)
1141 Most of the time, you run MPI code with SMPI to compute the time it
1142 would take to run it on a platform. But since the code is run through
1143 the ``smpirun`` script, you don't have any control on the launcher
1144 code, making it difficult to report the simulated time when the
1145 simulation ends. If you enable the ``smpi/display-timing`` item,
1146 ``smpirun`` will display this information when the simulation
1148 SMPI will also display information about the amout of real time spent
1149 in application code and in SMPI internals, to provide hints about the
1150 need to use sampling to reduce simulation time.
1152 .. _cfg=smpi/display-allocs:
1154 Reporting memory allocations
1155 ............................
1157 **Option** ``smpi/display-allocs`` **Default:** 0 (false)
1159 SMPI intercepts malloc and calloc calls performed inside the running
1160 application, if it wasn't compiled with SMPI_NO_OVERRIDE_MALLOC.
1161 With this option, SMPI will show at the end of execution the amount of
1162 memory allocated through these calls, and locate the most expensive one.
1163 This helps finding the targets for manual memory sharing, or the threshold
1164 to use for smpi/auto-shared-malloc-thresh option (see :ref:`cfg=smpi/auto-shared-malloc-thresh`).
1166 .. _cfg=smpi/keep-temps:
1168 Keeping temporary files after simulation
1169 ........................................
1171 **Option** ``smpi/keep-temps`` **default:** 0 (false)
1173 SMPI usually generates a lot of temporary files that are cleaned after
1174 use. This option requests to preserve them, for example to debug or
1175 profile your code. Indeed, the binary files are removed very early
1176 under the dlopen privatization schema, which tends to fool the
1179 .. _cfg=smpi/lat-factor:
1184 **Option** ``smpi/lat-factor`` |br|
1185 **default:** 65472:11.6436;15424:3.48845;9376:2.59299;5776:2.18796;3484:1.88101;1426:1.61075;732:1.9503;257:1.95341;0:2.01467
1187 The motivation and syntax for this option is identical to the motivation/syntax
1188 of :ref:`cfg=smpi/bw-factor`.
1190 There is an important difference, though: While smpi/bw-factor `reduces` the
1191 actual bandwidth (i.e., values between 0 and 1 are valid), latency factors
1192 increase the latency, i.e., values larger than or equal to 1 are valid here.
1194 .. _cfg=smpi/papi-events:
1196 Trace hardware counters with PAPI
1197 .................................
1199 **Option** ``smpi/papi-events`` **default:** unset
1201 When the PAPI support is compiled into SimGrid, this option takes the
1202 names of PAPI counters and adds their respective values to the trace
1203 files (See Section :ref:`tracing_tracing_options`).
1207 This feature currently requires superuser privileges, as registers
1208 are queried. Only use this feature with code you trust! Call
1209 smpirun for instance via ``smpirun -wrapper "sudo "
1210 <your-parameters>`` or run ``sudo sh -c "echo 0 >
1211 /proc/sys/kernel/perf_event_paranoid"`` In the later case, sudo
1212 will not be required.
1214 It is planned to make this feature available on a per-process (or per-thread?) basis.
1215 The first draft, however, just implements a "global" (i.e., for all processes) set
1216 of counters, the "default" set.
1218 .. code-block:: shell
1220 --cfg=smpi/papi-events:"default:PAPI_L3_LDM:PAPI_L2_LDM"
1222 .. _cfg=smpi/privatization:
1224 Automatic Privatization of Global Variables
1225 ...........................................
1227 **Option** ``smpi/privatization`` **default:** "dlopen" (when using smpirun)
1229 MPI executables are usually meant to be executed in separate
1230 processes, but SMPI is executed in only one process. Global variables
1231 from executables will be placed in the same memory region and shared
1232 between processes, causing intricate bugs. Several options are
1233 possible to avoid this, as described in the main `SMPI publication
1234 <https://hal.inria.fr/hal-01415484>`_ and in the :ref:`SMPI
1235 documentation <SMPI_what_globals>`. SimGrid provides two ways of
1236 automatically privatizing the globals, and this option allows one to
1237 choose between them.
1239 - **no** (default when not using smpirun): Do not automatically
1240 privatize variables. Pass ``-no-privatize`` to smpirun to disable
1242 - **dlopen** or **yes** (default when using smpirun): Link multiple
1243 times against the binary.
1244 - **mmap** (slower, but maybe somewhat more stable):
1245 Runtime automatic switching of the data segments.
1248 This configuration option cannot be set in your platform file. You can only
1249 pass it as an argument to smpirun.
1251 .. _cfg=smpi/privatize-libs:
1253 Automatic privatization of global variables inside external libraries
1254 .....................................................................
1256 **Option** ``smpi/privatize-libs`` **default:** unset
1258 **Linux/BSD only:** When using dlopen (default) privatization,
1259 privatize specific shared libraries with internal global variables, if
1260 they can't be linked statically. For example libgfortran is usually
1261 used for Fortran I/O and indexes in files can be mixed up.
1263 Multiple libraries can be given, semicolon separated.
1265 This configuration option can only use either full paths to libraries,
1266 or full names. Check with ldd the name of the library you want to
1269 .. code-block:: shell
1273 libgfortran.so.3 => /usr/lib/x86_64-linux-gnu/libgfortran.so.3 (0x00007fbb4d91b000)
1276 Then you can use ``--cfg=smpi/privatize-libs:libgfortran.so.3``
1277 or ``--cfg=smpi/privatize-libs:/usr/lib/x86_64-linux-gnu/libgfortran.so.3``,
1278 but not ``libgfortran`` nor ``libgfortran.so``.
1280 .. _cfg=smpi/send-is-detached-thresh:
1282 Simulating MPI detached send
1283 ............................
1285 **Option** ``smpi/send-is-detached-thresh`` **default:** 65536
1287 This threshold specifies the size in bytes under which the send will
1288 return immediately. This is different from the threshold detailed in
1289 :ref:`cfg=smpi/async-small-thresh` because the message is not
1290 really sent when the send is posted. SMPI still waits for the
1291 corresponding receive to be posted, in order to perform the communication
1294 .. _cfg=smpi/coll-selector:
1296 Simulating MPI collective algorithms
1297 ....................................
1299 **Option** ``smpi/coll-selector`` **Possible values:** naive (default), ompi, mpich
1301 SMPI implements more than 100 different algorithms for MPI collective
1302 communication, to accurately simulate the behavior of most of the
1303 existing MPI libraries. The ``smpi/coll-selector`` item can be used to
1304 select the decision logic either of the OpenMPI or the MPICH libraries. (By
1305 default SMPI uses naive version of collective operations.)
1307 Each collective operation can be manually selected with a
1308 ``smpi/collective_name:algo_name``. Available algorithms are listed in
1309 :ref:`SMPI_use_colls`.
1311 .. TODO:: All available collective algorithms will be made available
1312 via the ``smpirun --help-coll`` command.
1314 Add a barrier in MPI_Finalize
1315 .............................
1317 .. _cfg=smpi/finalization-barrier:
1319 **Option** ``smpi/finalization-barrier`` **default:** off
1321 By default, SMPI processes are destroyed as soon as soon as their code ends,
1322 so after a successful MPI_Finalize call returns. In some rare cases, some data
1323 might have been attached to MPI objects still active in the remaining processes,
1324 and can be destroyed eagerly by the finished process.
1325 If your code shows issues at finalization, such as segmentation fault, triggering
1326 this option will add an explicit MPI_Barrier(MPI_COMM_WORLD) call inside the
1327 MPI_Finalize, so that all processes will terminate at almost the same point.
1328 It might affect the total timing by the cost of a barrier.
1330 .. _cfg=smpi/errors-are-fatal:
1332 **Option** ``smpi/errors-are-fatal`` **default:** on
1334 By default, SMPI processes will crash if a MPI error code is returned. MPI allows
1335 to explicitely set MPI_ERRORS_RETURN errhandler to avoid this behaviour. This flag
1336 will turn on this behaviour by default (for all concerned types and errhandlers).
1337 This can ease debugging by going after the first reported error.
1339 .. _cfg=smpi/pedantic:
1341 **Option** ``smpi/pedantic`` **default:** on
1343 By default, SMPI will report all errors it finds in MPI codes. Some of these errors
1344 may not be considered as errors by all developers. This flag can be turned off to
1345 avoid reporting some usually harmless mistakes.
1346 Concerned errors list (will be expanded in the future):
1347 - Calling MPI_Win_fence only once in a program, hence just opening an epoch without
1350 .. _cfg=smpi/iprobe:
1352 Inject constant times for MPI_Iprobe
1353 ....................................
1355 **Option** ``smpi/iprobe`` **default:** 0.0001
1357 The behavior and motivation for this configuration option is identical
1358 with :ref:`smpi/test <cfg=smpi/test>`, but for the function
1361 .. _cfg=smpi/iprobe-cpu-usage:
1363 Reduce speed for iprobe calls
1364 .............................
1366 **Option** ``smpi/iprobe-cpu-usage`` **default:** 1 (no change)
1368 MPI_Iprobe calls can be heavily used in applications. To account
1369 correctly for the energy that cores spend probing, it is necessary to
1370 reduce the load that these calls cause inside SimGrid.
1372 For instance, we measured a maximum power consumption of 220 W for a
1373 particular application but only 180 W while this application was
1374 probing. Hence, the correct factor that should be passed to this
1375 option would be 180/220 = 0.81.
1379 Inject constant times for MPI_Init
1380 ..................................
1382 **Option** ``smpi/init`` **default:** 0
1384 The behavior and motivation for this configuration option is identical
1385 with :ref:`smpi/test <cfg=smpi/test>`, but for the function ``MPI_Init()``.
1389 Inject constant times for MPI_Isend()
1390 .....................................
1392 **Option** ``smpi/ois``
1394 The behavior and motivation for this configuration option is identical
1395 with :ref:`smpi/os <cfg=smpi/os>`, but for the function ``MPI_Isend()``.
1399 Inject constant times for MPI_send()
1400 ....................................
1402 **Option** ``smpi/os``
1404 In several network models such as LogP, send (MPI_Send, MPI_Isend) and
1405 receive (MPI_Recv) operations incur costs (i.e., they consume CPU
1406 time). SMPI can factor these costs in as well, but the user has to
1407 configure SMPI accordingly as these values may vary by machine. This
1408 can be done by using ``smpi/os`` for MPI_Send operations; for MPI_Isend
1409 and MPI_Recv, use ``smpi/ois`` and ``smpi/or``, respectively. These work
1410 exactly as ``smpi/ois``.
1412 This item can consist of multiple sections; each section takes three
1413 values, for example ``1:3:2;10:5:1``. The sections are divided by ";"
1414 so this example contains two sections. Furthermore, each section
1415 consists of three values.
1417 1. The first value denotes the minimum size in bytes for this section to take effect;
1418 read it as "if message size is greater than this value (and other section has a larger
1419 first value that is also smaller than the message size), use this".
1420 In the first section above, this value is "1".
1422 2. The second value is the startup time; this is a constant value that will always
1423 be charged, no matter what the size of the message. In the first section above,
1426 3. The third value is the `per-byte` cost. That is, it is charged for every
1427 byte of the message (incurring cost messageSize*cost_per_byte)
1428 and hence accounts also for larger messages. In the first
1429 section of the example above, this value is "2".
1431 Now, SMPI always checks which section it should use for a given
1432 message; that is, if a message of size 11 is sent with the
1433 configuration of the example above, only the second section will be
1434 used, not the first, as the first value of the second section is
1435 closer to the message size. Hence, when ``smpi/os=1:3:2;10:5:1``, a
1436 message of size 11 incurs the following cost inside MPI_Send:
1437 ``5+11*1`` because 5 is the startup cost and 1 is the cost per byte.
1439 Note that the order of sections can be arbitrary; they will be ordered internally.
1443 Inject constant times for MPI_Recv()
1444 ....................................
1446 **Option** ``smpi/or``
1448 The behavior and motivation for this configuration option is identical
1449 with :ref:`smpi/os <cfg=smpi/os>`, but for the function ``MPI_Recv()``.
1452 .. _cfg=smpi/grow-injected-times:
1454 Inject constant times for MPI_Test
1455 ..................................
1457 **Option** ``smpi/test`` **default:** 0.0001
1459 By setting this option, you can control the amount of time a process
1460 sleeps when MPI_Test() is called; this is important, because SimGrid
1461 normally only advances the time while communication is happening and
1462 thus, MPI_Test will not add to the time, resulting in deadlock if it is
1463 used as a break-condition as in the following example:
1468 MPI_Test(request, flag, status);
1472 To speed up execution, we use a counter to keep track of how often we
1473 checked if the handle is now valid or not. Hence, we actually
1474 use counter*SLEEP_TIME, that is, the time MPI_Test() causes the
1475 process to sleep increases linearly with the number of previously
1476 failed tests. This behavior can be disabled by setting
1477 ``smpi/grow-injected-times`` to **no**. This will also disable this
1478 behavior for MPI_Iprobe.
1480 .. _cfg=smpi/shared-malloc:
1481 .. _cfg=smpi/shared-malloc-hugepage:
1486 **Option** ``smpi/shared-malloc`` **Possible values:** global (default), local
1488 If your simulation consumes too much memory, you may want to modify
1489 your code so that the working areas are shared by all MPI ranks. For
1490 example, in a block-cyclic matrix multiplication, you will only
1491 allocate one set of blocks, and all processes will share them.
1492 Naturally, this will lead to very wrong results, but this will save a
1493 lot of memory. So this is still desirable for some studies. For more on
1494 the motivation for that feature, please refer to the `relevant section
1495 <https://simgrid.github.io/SMPI_CourseWare/topic_understanding_performance/matrixmultiplication>`_
1496 of the SMPI CourseWare (see Activity #2.2 of the pointed
1497 assignment). In practice, change the calls for malloc() and free() into
1498 SMPI_SHARED_MALLOC() and SMPI_SHARED_FREE().
1500 SMPI provides two algorithms for this feature. The first one, called
1501 ``local``, allocates one block per call to SMPI_SHARED_MALLOC()
1502 (each call site gets its own block) ,and this block is shared
1503 among all MPI ranks. This is implemented with the shm_* functions
1504 to create a new POSIX shared memory object (kept in RAM, in /dev/shm)
1505 for each shared block.
1507 With the ``global`` algorithm, each call to SMPI_SHARED_MALLOC()
1508 returns a new address, but it only points to a shadow block: its memory
1509 area is mapped on a 1 MiB file on disk. If the returned block is of size
1510 N MiB, then the same file is mapped N times to cover the whole block.
1511 At the end, no matter how many times you call SMPI_SHARED_MALLOC, this will
1512 only consume 1 MiB in memory.
1514 You can disable this behavior and come back to regular mallocs (for
1515 example for debugging purposes) using ``no`` as a value.
1517 If you want to keep private some parts of the buffer, for instance if these
1518 parts are used by the application logic and should not be corrupted, you
1519 can use SMPI_PARTIAL_SHARED_MALLOC(size, offsets, offsets_count). For example:
1523 mem = SMPI_PARTIAL_SHARED_MALLOC(500, {27,42 , 100,200}, 2);
1525 This will allocate 500 bytes to mem, such that mem[27..41] and
1526 mem[100..199] are shared while other area remain private.
1528 Then, it can be deallocated by calling SMPI_SHARED_FREE(mem).
1530 When smpi/shared-malloc:global is used, the memory consumption problem
1531 is solved, but it may induce too much load on the kernel's pages table.
1532 In this case, you should use huge pages so that the kernel creates only one
1533 entry per MB of malloced data instead of one entry per 4 kB.
1534 To activate this, you must mount a hugetlbfs on your system and allocate
1535 at least one huge page:
1537 .. code-block:: shell
1540 sudo mount none /home/huge -t hugetlbfs -o rw,mode=0777
1541 sudo sh -c 'echo 1 > /proc/sys/vm/nr_hugepages' # echo more if you need more
1543 Then, you can pass the option
1544 ``--cfg=smpi/shared-malloc-hugepage:/home/huge`` to smpirun to
1545 actually activate the huge page support in shared mallocs.
1547 .. _cfg=smpi/auto-shared-malloc-thresh:
1549 Automatically share allocations
1550 ...............................
1552 **Option** ``smpi/auto-shared-malloc-thresh:`` **Default:** 0 (false)
1553 This value in bytes represents the size above which all allocations
1554 will be "shared" by default (as if they were performed through
1555 SMPI_SHARED_MALLOC macros). Default = 0 = disabled feature.
1556 The value must be carefully chosen to only select data buffers which
1557 will not modify execution path or cause crash if their content is false.
1558 Option :ref:`cfg=smpi/display-allocs` can be used to locate the largest
1559 allocation detected in a run, and provide a good starting threshold.
1560 Note : malloc, calloc and free are overridden by smpicc/cxx by default.
1561 This can cause some troubles if codes are already overriding these. If this
1562 is the case, defining SMPI_NO_OVERRIDE_MALLOC in the compilation flags can
1563 help, but will make this feature unusable.
1567 Inject constant times for MPI_Wtime, gettimeofday and clock_gettime
1568 ...................................................................
1570 **Option** ``smpi/wtime`` **default:** 10 ns
1572 This option controls the amount of (simulated) time spent in calls to
1573 MPI_Wtime(), gettimeofday() and clock_gettime(). If you set this value
1574 to 0, the simulated clock is not advanced in these calls, which leads
1575 to issues if your application contains such a loop:
1579 while(MPI_Wtime() < some_time_bound) {
1580 /* some tests, with no communication nor computation */
1583 When the option smpi/wtime is set to 0, the time advances only on
1584 communications and computations. So the previous code results in an
1585 infinite loop: the current [simulated] time will never reach
1586 ``some_time_bound``. This infinite loop is avoided when that option
1587 is set to a small value, as it is by default since SimGrid v3.21.
1589 Note that if your application does not contain any loop depending on
1590 the current time only, then setting this option to a non-zero value
1591 will slow down your simulations by a tiny bit: the simulation loop has
1592 to be broken out of and reset each time your code asks for the current time.
1593 If the simulation speed really matters to you, you can avoid this
1594 extra delay by setting smpi/wtime to 0.
1596 .. _cfg=smpi/list-leaks:
1598 Report leaked MPI objects
1599 .........................
1601 **Option** ``smpi/list-leaks`` **default:** 0
1603 This option controls whether to report leaked MPI objects.
1604 The parameter is the number of leaks to report.
1606 Other Configurations
1607 --------------------
1609 .. _cfg=debug/clean-atexit:
1611 Cleanup at Termination
1612 ......................
1614 **Option** ``debug/clean-atexit`` **default:** on
1616 If your code is segfaulting during its finalization, it may help to
1617 disable this option to request that SimGrid not attempt any cleanups at
1618 the end of the simulation. Since the Unix process is ending anyway,
1619 the operating system will wipe it all.
1626 **Option** ``path`` **default:** . (current dir)
1628 It is possible to specify a list of directories to search in for the
1629 trace files (see :ref:`pf_trace`) by using this configuration
1630 item. To add several directory to the path, set the configuration
1631 item several times, as in ``--cfg=path:toto --cfg=path:tutu``
1633 .. _cfg=debug/breakpoint:
1638 **Option** ``debug/breakpoint`` **default:** unset
1640 This configuration option sets a breakpoint: when the simulated clock
1641 reaches the given time, a SIGTRAP is raised. This can be used to stop
1642 the execution and get a backtrace with a debugger.
1644 It is also possible to set the breakpoint from inside the debugger, by
1645 writing in global variable simgrid::simix::breakpoint. For example,
1648 .. code-block:: shell
1650 set variable simgrid::simix::breakpoint = 3.1416
1652 .. _cfg=debug/verbose-exit:
1657 **Option** ``debug/verbose-exit`` **default:** on
1659 By default, when Ctrl-C is pressed, the status of all existing actors
1660 is displayed before exiting the simulation. This is very useful to
1661 debug your code, but it can become troublesome if you have many
1662 actors. Set this configuration item to **off** to disable this
1665 .. _cfg=exception/cutpath:
1667 Truncate local path from exception backtrace
1668 ............................................
1670 **Option** ``exception/cutpath`` **default:** off
1672 This configuration option is used to remove the path from the
1673 backtrace shown when an exception is thrown. This is mainly useful for
1674 the tests: the full file path would makes the tests non-reproducible because
1675 the paths of source files depend of the build settings. That would
1676 break most of the tests since their output is continually compared.
1680 Logging configuration
1681 ---------------------
1683 As introduced in :ref:`outcome_logs`, the SimGrid logging mechanism allows to configure at runtime the messages that should be displayed and those that should be omitted. Each
1684 message produced in the code is given a category (denoting its topic) and a priority. Then at runtime, each category is given a threshold (only messages of priority higher than
1685 that threshold are displayed), a layout (deciding how the messages in this category are formatted), and an appender (deciding what to do with the message: either print on stderr or
1688 This section explains how to configure this logging features. You can also refer to the documentation of the :ref:`programmer's interface <logging_prog>`, that allows to produce
1689 messages from your code.
1691 Most of the time, the logging mechanism is configured at runtime using the ``--log`` command-line argument, even if you can also use :c:func:`xbt_log_control_set()` to control it from
1692 your program. To pass configure more than one setting, you can either pass several ``--log`` arguments, or separate your settings with spaces, that must be quoted accordingly. In
1693 practice, the following is equivalent to the above settings: ``--log=root.thresh:error --log=s4u_host.thresh:debug``.
1695 If you want to specify more than one setting, you can either pass several ``--log`` argument to your program as above, or separate them with spaces. In this case, you want to quote
1696 your settings, as in ``--log="root.thresh:error s4u_host.thresh:debug"``. The parameters are interpreted in order, from left to right.
1699 Threshold configuration
1700 .......................
1702 The keyword ``threshold`` controls which logging event will get displayed in a given category. For example, ``--log=root.threshold:debug`` displays *every* message produced in the
1703 ``root`` category and its subcategories (i.e., every message produced -- this is *extremely* verbose), while ``--log=root.thres:critical`` turns almost everything off. As you can
1704 see, ``threshold`` can be abbreviated here.
1706 Existing thresholds:
1708 - ``trace`` some functions display a message at this level when entering or returning
1709 - ``debug`` output that is mostly useful when debugging the corresponding module.
1710 - ``verbose`` verbose output that is only mildly interesting and can easily be ignored
1711 - ``info`` usual output (this is the default threshold of all categories)
1712 - ``warning`` minor issue encountered
1713 - ``error`` issue encountered
1714 - ``critical`` major issue encountered, such as assertions failures
1718 Format configuration
1719 ....................
1721 The keyword ``fmt`` controls the layout (the format) of a logging category. For example, ``--log=root.fmt:%m`` reduces the output to the user-message only, removing any decoration such
1722 as the date, or the actor ID, everything. Existing format directives:
1725 - %n: line separator (LOG4J compatible)
1726 - %e: plain old space (SimGrid extension)
1728 - %m: user-provided message
1730 - %c: Category name (LOG4J compatible)
1731 - %p: Priority name (LOG4J compatible)
1733 - %h: Hostname (SimGrid extension)
1734 - %a: Actor name (SimGrid extension -- note that with SMPI this is the integer value of the process rank)
1735 - %i: Actor PID (SimGrid extension -- this is a 'i' as in 'i'dea)
1736 - %t: Thread "name" (LOG4J compatible -- actually the address of the thread in memory)
1738 - %F: file name where the log event was raised (LOG4J compatible)
1739 - %l: location where the log event was raised (LOG4J compatible, like '%%F:%%L' -- this is a l as in 'l'etter)
1740 - %L: line number where the log event was raised (LOG4J compatible)
1741 - %M: function name (LOG4J compatible -- called method name here of course).
1743 - %d: date (UNIX-like epoch)
1744 - %r: application age (time elapsed since the beginning of the application)
1747 ``--log=root.fmt:'[%h:%a:(%i) %r] %l: %m%n'`` gives you the default layout used for info messages while ``--log=root.fmt:'[%h:%a:(%i) %r] %l: [%c/%p] %m%n'`` gives you the default
1748 layout for the other priorities (it adds the source code location). Also, the actor identification is omitted by the default layout for the messages coming directly from the
1749 SimGrid kernel, so info messages are formatted with ``[%r] [%c/%p] %m%n`` in this case. When specifying the layout manually, such distinctions are currently impossible, and the
1750 provided layout is used for every messages.
1752 As with printf, you can specify the precision and width of the fields. For example, ``%.4r`` limits the date precision to four digits while ``%15h`` limits the host name to at most
1756 If you want to have spaces in your log format, you should protect it. Otherwise, SimGrid will consider that this is a space-separated list of several parameters. But you should
1757 also protect it from the shell that also splits command line arguments on spaces. At the end, you should use something such as ``--log="'root.fmt:%l: [%p/%c]: %m%n'"``.
1758 Another option is to use the ``%e`` directive for spaces, as in ``--log=root.fmt:%l:%e[%p/%c]:%e%m%n``.
1763 The keyword ``app`` controls the appended of a logging category. For example ``--log=root.app:file:mylogfile`` redirects every output to the file ``mylogfile``.
1765 With the ``splitfile`` appender, a new file is created when the size of the output reaches the specified size. The format is ``--log=root.app:splitfile:<size>:<file name>``. For
1766 example, ``--log=root.app:splitfile:500:mylog_%`` creates log files of at most 500 bytes, using the names ``mylog_0``, ``mylog_1``, ``mylog_2``, etc.
1768 The ``rollfile`` appender uses one file only, but the file is emptied and recreated when its size reaches the specified maximum. For example, ``--log=root.app:rollfile:500:mylog``
1769 ensures that the log file ``mylog`` will never overpass 500 bytes in size.
1771 Any appender setup this way have its own layout format, that you may change afterward. When specifying a new appender, its additivity is set to false to prevent log event displayed
1772 by this appender to "leak" to any other appender higher in the hierarchy. You can naturally change that if you want your messages to be displayed twice.
1777 The keyword ``add`` controls the additivity of a logging category. By default, the messages are only passed one appender only: the more specific, i.e. the first one found when
1778 climbing the tree from the category in which they were produced. In Log4J parlance, it is said that the default additivity of appenders is false. If you change this setting to
1779 ``on`` (or ``yes`` or ``1``), the produced messages will also be passed to the upper appender.
1781 Let's consider a more complex example: ``--log="root.app:file:all.log s4u.app:file:iface.log xbt.app:file:xbt.log xbt.add:yes``. Here, the logging of s4u will be sent to the
1782 ``iface.log`` file; the logging of the xbt toolbox will be sent to both the ``xbt.log`` file and the ``all.log`` file (because xbt additivity was enabled); and every other loggings
1783 will only be sent to ``all.log``.
1788 ``--help-logs`` displays a complete help message about logging in SimGrid.
1790 ``--help-log-categories`` displays the actual hierarchy of log categories for this binary.
1792 ``--log=no_loc`` hides the source locations (file names and line numbers) from the messages. This is useful to make tests reproducible.